M3-Verse

# $M^3$-Verse Benchmark This repository contains the $M^3$-Verse benchmark, a comprehensive benchmark for 3D and video Question Answering (QA) with multi-state observations. The dataset is designed to test agents' capabilities in spatial understanding, temporal reasoning, attribute recognition, and hallucination detection within dynamic 3D environments. The $M^3$-Verse benchmark consists of two distinct subsets: 1. **$M^3$-Verse-Sim**: The simulation subset, featuring large-scale synthetic data with rich ground-truth metadata. 2. **$M^3$-Verse-Real**: The real-world subset, featuring video data captured in real physical environments. Please refer to the README files in each subdirectory for specific details about the data structure and usage of each subset. - [$M^3$-Verse-Sim Documentation](./M3-Verse-Sim/README.md) - [$M^3$-Verse-Real Documentation](./M3-Verse-Real/README.md)